Fi
n
Ll
a
m
a
: Fi
nan
ci
a
l S
en
t
im
en
t
Cl
a
ssi
f
ic
a
t
i
on
f
or
Alg
or
i
t
h
mic T
rad
i
n
g A
pp
lic
a
t
i
on
s
T
hano
s K
on
s
t
an
t
i
n
i
d
is
1
,
G
i
o
rg
o
s I
a
c
o
v
i
de
s
1
,
2
, Mi
n
g
x
ue
X
u
1
, T
on
y
G
. C
on
s
t
an
t
i
n
i
de
s
1
,
D
an
il
o
M
and
ic
1
1
Im
pe
ri
a
l C
o
ll
e
g
e
L
ondon
,
2
MIT
{
a
.
k
on
s
t
an
t
i
n
i
d
is16, gi
o
rg
o
s.i
a
c
o
v
i
de
s20, m.
x
u
21,
a
.c
on
s
t
an
t
i
n
i
de
s,
d
.m
and
ic}@im
pe
ri
a
l.
a
c.
u
k
A
b
s
t
r
ac
t
—
Th
e
r
e
ar
e
multipl
e
s
o
urc
e
s
o
f
f
inancial n
e
ws
o
nlin
e
which in
f
lu
e
nc
e
mark
e
t m
o
v
e
m
e
nts and trad
e
r’s d
e
cisi
o
ns. This
highlights th
e
n
ee
d
f
o
r accurat
e
s
e
ntim
e
nt anal
y
sis, in additi
o
n t
o
ha
v
ing appr
o
priat
e
alg
o
rithmic trading t
e
chniqu
e
s, t
o
arri
v
e
at
b
e
tt
e
r in
f
o
rm
e
d trading d
e
cisi
o
ns. Standard l
e
x
ic
o
n bas
e
d
s
e
ntim
e
nt appr
o
ach
e
s ha
v
e
d
e
m
o
nstrat
e
d th
e
ir p
o
w
e
r in aiding
f
inancial d
e
cisi
o
ns. H
o
w
e
v
e
r, th
e
y
ar
e
kn
o
wn t
o
su
ff
e
r
f
r
o
m issu
e
s
r
e
lat
e
d t
o
c
o
nt
e
x
t s
e
nsiti
v
it
y
and w
o
rd
o
rd
e
ring. Larg
e
Languag
e
M
o
d
e
ls
(
LL
M
s
)
can als
o
b
e
us
e
d in this c
o
nt
e
x
t, but th
e
y
ar
e
n
o
t
f
inanc
e
-
sp
e
ci
f
ic and t
e
nd t
o
r
e
quir
e
signi
f
icant c
o
mputati
o
nal
r
e
s
o
urc
e
s. T
o
f
acilitat
e
a
f
inanc
e
sp
e
ci
f
ic LL
M
f
ram
e
w
o
rk, w
e
intr
o
duc
e
a n
o
v
e
l appr
o
ach bas
e
d
o
n th
e
Llama 2 7B
f
o
undati
o
nal
m
o
d
e
l, in
o
rd
e
r t
o
b
e
n
ef
it
f
r
o
m its g
e
n
e
rati
v
e
natur
e
and c
o
mpr
e
-
h
e
nsi
v
e
languag
e
manipulati
o
n. This is achi
e
v
e
d b
y
f
in
e
-
tuning
th
e
Llama2 7B m
o
d
e
l
o
n a small p
o
rti
o
n
o
f
sup
e
r
v
is
e
d
f
inancial
s
e
ntim
e
nt anal
y
sis data, s
o
as t
o
j
o
intl
y
handl
e
th
e
c
o
mpl
e
x
iti
e
s
o
f
f
inancial l
e
x
ic
o
n and c
o
nt
e
x
t, and
f
urth
e
r
e
quipping it with a
n
e
ural n
e
tw
o
rk bas
e
d d
e
cisi
o
n m
e
chanism. Such a g
e
n
e
rat
o
r
-
classi
f
i
e
r sch
e
m
e
, r
e
f
e
rr
e
d t
o
as FinLlama, is train
e
d n
o
t
o
nl
y
t
o
classi
fy
th
e
s
e
ntim
e
nt
v
al
e
nc
e
but als
o
quanti
fy
its str
e
ngth, thus
o
ff
e
ring trad
e
rs a nuanc
e
d insight int
o
f
inancial n
e
ws articl
e
s.
C
o
mpl
e
m
e
nting this, th
e
impl
e
m
e
ntati
o
n
o
f
param
e
t
e
r
-
e
f
f
ici
e
nt
f
in
e
-
tuning thr
o
ugh L
o
RA
o
ptimis
e
s trainabl
e
param
e
t
e
rs, thus
minimising c
o
mputati
o
nal and m
e
m
o
r
y
r
e
quir
e
m
e
nts, with
o
ut
sacri
f
icing accurac
y
. Simulati
o
n r
e
sults d
e
m
o
nstrat
e
th
e
abilit
y
o
f
th
e
pr
o
p
o
s
e
d FinLlama t
o
pr
o
v
id
e
a
f
ram
e
w
o
rk
f
o
r
e
nhanc
e
d
p
o
rt
f
o
li
o
manag
e
m
e
nt d
e
cisi
o
ns and incr
e
as
e
d mark
e
t r
e
turns.
Th
e
s
e
r
e
sults und
e
rpin th
e
abilit
y
o
f
FinLlama t
o
c
o
nstruct high
-
r
e
turn p
o
rt
f
o
li
o
s which
e
x
hibit
e
nhanc
e
d r
e
sili
e
nc
e
,
e
v
e
n during
v
o
latil
e
p
e
ri
o
ds and unpr
e
dictabl
e
mark
e
t
e
v
e
nts.
In
de
x
T
e
r
ms
-
L
a
rg
e
l
an
g
ua
g
e
m
ode
ls,
f
i
nan
c
e
, s
en
t
im
en
t
ana
l
y
sis,
a
lg
o
ri
t
h
mic
t
r
ad
i
n
g,
pa
r
a
m
e
t
e
r
-
e
f
f
ici
en
t
f
i
ne
-
t
un
i
n
g
I. I
N
TR
O
D
U
CTI
ON
T
he
e
v
e
r i
n
cr
ea
si
n
g
p
r
o
mi
nen
c
e
o
f
a
lg
o
ri
t
h
mic
t
r
ad
i
n
g i
n
quan
t
i
t
a
t
i
v
e
f
i
nan
c
e
ha
s
ne
c
e
ssi
t
a
t
ed
t
he
need
f
o
r r
e
li
ab
l
e
and
a
c
t
i
onab
l
e
AI
-
a
i
ded
i
n
t
e
llig
en
c
e
f
r
o
m
v
a
s
t
s
t
r
ea
ms
o
f
da
t
a
wi
t
h
m
u
l
t
i
p
l
e
m
oda
li
t
i
e
s.
Of
pa
r
t
ic
u
l
a
r i
n
t
e
r
e
s
t
is g
ene
r
a
t
i
v
e
AI,
o
w
-
i
n
g
t
o
i
t
s
ab
ili
t
y
t
o
d
is
t
ill i
n
sig
h
t
s
f
r
o
m
non
-
nu
m
e
ric
a
l s
ou
rc
e
s
s
u
c
h
a
s
ne
ws,
ea
r
n
i
n
gs c
a
lls,
f
i
nan
ci
a
l r
epo
r
t
s,
and
o
t
he
r
t
e
x
t
ua
l
s
ou
rc
e
s. I
n
t
h
is c
on
t
e
x
t
, s
en
t
im
en
t
ana
l
y
sis
p
r
o
mis
e
s
t
o
b
ri
d
g
e
t
he
g
ap
be
t
w
een
m
a
r
k
e
t
m
o
v
e
m
en
t
s c
au
s
ed
b
y
g
eopo
li
t
ic
a
l
and
s
o
ci
oe
c
ono
mic
e
v
en
t
s,
hu
m
an
a
c
t
i
on
s,
and
quan
t
i
t
a
t
i
v
e
t
r
ad
i
n
g.
S
en
t
im
en
t
ana
l
y
sis r
e
s
t
s
upon
t
he
quan
t
i
f
ic
a
t
i
on
o
f
op
i
n
i
on
s
p
r
e
s
en
t
i
n
un
l
abe
l
ed
t
e
x
t
ua
l
da
t
a
,
and
a
ims
t
o
c
a
t
e
g
o
ri
z
e
w
he
t
he
r
t
he
o
v
e
r
a
ll
v
i
e
w
po
i
n
t
is
po
si
t
i
v
e
,
ne
g
a
t
i
v
e
,
o
r
neu
t
r
a
l.
W
hen
app
li
ed
t
o
l
a
rg
e
-
sc
a
l
e
i
n
f
o
rm
a
t
i
on
s
ou
rc
e
s,
t
h
is m
a
y
y
i
e
l
d
an
unde
rs
t
and
i
n
g
o
f
t
he
o
v
e
r
a
ll
d
ir
e
c
t
i
on
o
f
m
a
cr
o
sc
op
ic
t
r
end
s,
a
t
a
s
k
w
h
ic
h
is
bo
t
h
c
ha
ll
en
gi
n
g
and
t
im
e
-
c
on
s
u
mi
n
g
f
o
r
hu
m
an
ana
l
y
s
t
s. Im
po
r
t
an
t
l
y
,
t
he
s
en
t
im
en
t
c
on
t
a
i
ned
i
n
on
-
li
ne
t
e
x
t
ua
l s
ou
rc
e
s c
an
d
ri
v
e
m
a
r
k
e
t
m
o
v
e
m
en
t
s; s
u
c
h
i
n
f
o
rm
a
t
i
on
ha
r
bou
rs i
n
t
ri
n
sic
ad
v
an
t
a
g
e
s
and
a
c
o
m
pe
t
i
t
i
v
e
ed
g
e
t
o
t
ho
s
e
equ
i
pped
wi
t
h
t
he
t
oo
ls
t
o
ha
r
ne
ss i
t
.
D
e
s
p
i
t
e
c
on
c
ep
t
ua
l
benef
i
t
s,
t
he
he
t
e
r
o
g
eneou
s,
nuan
c
ed
,
and
v
a
s
t
na
t
u
r
e
o
f
f
i
nan
ci
a
l
t
e
x
t
p
r
e
s
en
t
s
un
i
que
c
ha
ll
en
g
e
s
w
hen
i
t
c
o
m
e
s
t
o
e
x
t
r
a
c
t
i
n
g s
en
t
im
en
t
i
n
a
m
anne
r
t
ha
t
is
bo
t
h
a
cc
u
r
a
t
e
and
a
c
t
i
onab
l
e
. F
o
r
e
x
a
m
p
l
e
,
t
he
w
o
r
d
s ‘
bu
ll’
and
‘
bea
r’
a
r
e
neu
t
r
a
l i
n
t
he
g
ene
r
a
l
v
o
c
abu
l
a
r
y
,
bu
t
i
n
f
i
nan
ci
a
l
m
a
r
k
e
t
s,
t
he
ir r
e
s
pe
c
t
i
v
e
c
onno
t
a
t
i
on
s
a
r
e
s
t
ric
t
l
y
po
si
t
i
v
e
o
r
ne
g
a
t
i
v
e
[
1
]
. T
h
is
h
ig
h
lig
h
t
s
t
he
need
f
o
r c
on
t
e
x
t
-
a
w
a
r
e
s
en
t
im
en
t
e
x
t
r
a
c
t
i
on
,
and
unde
r
p
i
n
s
t
he
c
o
m
p
l
e
x
i
t
i
e
s
o
f
na
t
u
r
a
l
l
an
g
ua
g
e
p
r
o
c
e
ssi
n
g
(N
L
P)
i
n
f
i
nan
ci
a
l
app
lic
a
t
i
on
s.
T
o
ad
r
e
ss
t
he
s
e
iss
ue
s, w
e
c
on
si
de
r
a
t
w
o
-f
o
l
d
f
unda
m
en
t
a
l
que
s
t
i
on
:
•
C
an
l
a
rg
e
l
an
g
ua
g
e
m
ode
ls
(
LLMs
)
, w
h
ic
h
ha
v
e
r
e
v
o
l
u
-
t
i
on
i
z
ed
m
an
i
f
o
l
d
a
r
ea
s
o
f
N
L
P
,
be
s
pe
ci
f
ic
a
ll
y
t
a
il
o
r
ed
f
o
r s
en
t
im
en
t
ana
l
y
sis i
n
t
he
f
i
nan
c
e
do
m
a
i
n
,
pa
r
t
ic
u
l
a
rl
y
f
o
r
a
lg
o
ri
t
h
mic
t
r
ad
i
n
g?
•
C
an
t
h
is
t
a
il
o
ri
n
g
be
a
c
h
i
e
v
ed
i
n
a
w
a
y
w
h
ic
h
doe
s
no
t
r
equ
ir
e
t
he
v
a
s
t
c
o
m
pu
t
a
t
i
ona
l r
e
s
ou
rc
e
s,
t
y
p
ic
a
ll
y
a
ss
o
ci
a
t
ed
wi
t
h
N
L
P
m
ode
ls,
t
hu
s m
a
k
i
n
g
t
he
app
r
oa
c
h
a
cc
e
ssi
b
l
e
t
o
a
b
r
oade
r
aud
i
en
c
e
equ
i
pped
wi
t
h
s
t
anda
r
d
c
o
m
pu
t
a
t
i
ona
l r
e
s
ou
rc
e
s?
.
O
u
r
p
r
opo
s
ed
Fi
n
Lla
m
a is
one
s
u
c
h
s
o
l
u
t
i
on
, w
h
ic
h
is
ob
t
a
i
ned
b
y
f
i
ne
-
t
un
i
n
g
a
p
r
e
-
t
r
a
i
ned
LLM
(
na
m
e
l
y
Ll
a
m
a
2
7B
[
2
])
on
s
pe
ci
a
lis
ed
, l
abe
ll
ed
and
pub
licl
y
a
v
a
il
ab
l
e
f
i
nan
ci
a
l
ne
ws
da
t
a
s
e
t
s. T
he
u
l
t
im
a
t
e
g
oa
l
o
f
Fi
n
Ll
a
m
a
is
t
o
enhan
c
e
t
he
pe
r
f
o
rm
an
c
e
o
f
f
i
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis, w
h
ils
t
l
e
v
e
r
a
gi
n
g
on
pa
r
a
m
e
t
e
r
-
e
f
f
ici
en
t
f
i
ne
-
t
un
i
n
g
(P
EFT
)
and
8
-
b
i
t
quan
t
i
z
a
-
t
i
on
,
t
h
r
ou
g
h
L
o
RA
[
3
]
,
t
o
mi
n
imis
e
r
e
s
ou
rc
e
u
t
ilis
a
t
i
on
.
T
he
m
a
i
n
c
on
t
ri
bu
t
i
on
s
o
f
t
h
is w
o
r
k
a
r
e
:
•
Targ
e
t
e
d
f
in
e
-
tuning: R
a
t
he
r
t
han
u
t
ilisi
n
g
one
o
v
e
r
-
a
rc
h
i
n
g m
ode
l
f
o
r
d
i
v
e
rs
e
f
i
nan
ci
a
l
t
a
s
k
s,
ou
r
app
r
oa
c
h
c
ap
i
t
a
li
z
e
s
on
t
he
f
ounda
t
i
ona
l
p
r
e
-
t
r
a
i
ned
Ll
a
m
a
2
m
ode
l, w
he
r
eb
y
f
i
ne
-
t
un
i
n
g is
pe
r
f
o
rm
ed
s
pe
ci
f
ic
a
ll
y
f
o
r
t
he
pu
r
po
s
e
o
f
s
en
t
im
en
t
cl
a
ssi
f
ic
a
t
i
on
t
h
r
ou
g
h
a
S
o
f
t
M
a
x
cl
a
ssi
f
ic
a
t
i
on
l
a
y
e
r
a
t
i
t
s
ou
t
pu
t
.
•
E
f
f
ici
e
nt r
e
s
o
urc
e
utili
z
ati
o
n:
O
u
r
app
r
oa
c
h
en
s
u
r
e
s
t
ha
t
e
v
en
s
t
anda
r
d
c
o
m
pu
t
a
t
i
ona
l r
e
s
ou
rc
e
s, wi
t
h
no
h
ig
h
-
end
GPU
s, c
an
be
e
m
p
l
o
y
ed
. B
y
v
ir
t
ue
o
f
t
he
p
r
e
-
t
r
a
i
ned
Ll
a
m
a
2 m
ode
l
and
t
h
r
ou
g
h
t
a
rg
e
t
ed
pa
r
a
m
e
t
e
r
-
e
f
f
ici
en
t
f
i
ne
-
t
un
i
n
g, w
e
d
r
a
m
a
t
ic
a
ll
y
r
edu
c
e
c
o
m
pu
t
a
t
i
ona
l
de
-
m
and
s c
o
m
pa
r
ed
t
o
t
he
e
x
is
t
i
n
g m
e
t
hod
s,
t
hu
s
b
ri
d
gi
n
g
1
arXiv:2403.12285v1
[cs.CL]
18 Mar 2024
t
he
g
ap
be
t
w
een
a
c
ade
mic
ben
c
h
m
a
r
k
s
and
p
r
a
c
t
ic
a
l
u
t
ili
t
y
.
•
B
e
nchmarking and r
e
al
-
w
o
rld applicati
o
n: T
he
s
u
cc
e
ss
o
f
f
i
ne
-
t
uned
LLMs
f
o
r
f
i
nan
c
e
ha
s
a
ls
o
h
ig
h
lig
h
t
ed
t
ha
t
t
he
do
m
a
i
n
o
f
po
r
t
f
o
li
o
c
on
s
t
r
u
c
t
i
on
ha
s
no
t
y
e
t
been
adequa
t
e
l
y
add
r
e
ss
ed
. T
o
t
h
is
end
, w
e
i
n
t
e
gr
a
t
e
t
he
e
x
t
r
a
c
t
ed
s
en
t
im
en
t
sig
na
ls
b
y
Fi
n
Ll
a
m
a
i
n
t
o
a
l
on
g
-
s
ho
r
t
po
r
t
f
o
li
o
, w
h
ic
h
a
ll
o
ws
u
s
t
o
ob
t
a
i
n
f
i
nan
c
e
-
s
pe
ci
f
ic
r
ea
l
-
w
o
rl
d
m
e
t
rics i
n
cl
ud
i
n
g c
u
m
u
l
a
t
i
v
e
r
e
t
u
r
n
s
and
t
he
S
ha
r
pe
r
a
t
i
o
.
II. R
ELATE
D
W
O
RK
T
he
po
t
en
t
i
a
l
o
f
s
en
t
im
en
t
ana
l
y
sis i
n
f
i
nan
c
e
w
a
s
f
irs
t
r
e
c
o
g
n
is
ed
b
y
F
a
m
a
w
ho
, i
n
1970, i
n
t
r
odu
c
ed
t
he
no
t
i
on
o
f
t
he
E
f
f
ici
en
t
M
a
r
k
e
t
H
y
po
t
he
sis
(
EM
H
)
[
4
]
, w
h
ic
h
s
t
a
t
e
s
t
ha
t
s
t
o
c
k
p
ric
e
s c
han
g
e
i
n
r
e
s
pon
s
e
t
o
une
x
pe
c
t
ed
f
unda
m
en
t
a
l
i
n
f
o
rm
a
t
i
on
and
ne
ws. I
n
t
h
is c
on
t
e
x
t
,
be
f
o
r
e
t
he
i
n
t
r
odu
c
t
i
on
o
f
ad
v
an
c
ed
m
a
c
h
i
ne
l
ea
r
n
i
n
g
t
oo
ls,
t
he
f
i
nan
ci
a
l s
e
c
t
o
r
e
m
-
p
l
o
y
ed
l
e
x
ic
on
-
d
ri
v
en
app
r
oa
c
he
s
[
1
]
. T
he
s
e
m
e
t
hod
s
ana
l
y
s
e
t
e
x
t
ua
l c
on
t
en
t
, s
ou
rc
ed
f
r
o
m
ne
ws
a
r
t
icl
e
s
o
r
f
i
nan
ci
a
l
d
iscl
o
-
s
u
r
e
s,
ba
s
ed
on
s
pe
ci
f
ic
k
e
y
w
o
r
d
s, w
h
ic
h
a
r
e
t
hen
li
n
k
ed
t
o
e
s
t
ab
lis
hed
s
en
t
im
en
t
r
a
t
i
n
gs
[
5
]
,
[
6
]
.
H
o
w
e
v
e
r,
an
e
x
ponen
t
i
a
l
i
n
cr
ea
s
e
i
n
t
he
v
o
l
u
m
e
and
ric
hne
ss
o
f
a
v
a
il
ab
l
e
i
n
f
o
rm
a
t
i
on
ha
s
opened
a
f
e
r
t
il
e
gr
ound
f
o
r m
a
c
h
i
ne
l
ea
r
n
i
n
g, i
n
cl
ud
i
n
g
t
e
c
hn
i
que
s s
u
c
h
a
s
N
a
i
v
e
B
a
y
e
s
and
S
uppo
r
t
V
e
c
t
o
r M
a
c
h
i
ne
s
[
7
]
, w
h
ic
h
a
r
e
s
u
mm
a
ris
ed
i
n
Fig
u
r
e
1.
I
n
pa
r
a
ll
e
l,
t
he
ad
v
an
c
e
s i
n
deep
l
ea
r
n
i
n
g
ha
v
e
be
c
o
m
e
i
n
s
t
r
u
m
en
t
a
l
f
o
r
N
L
P
r
e
s
ea
rc
h
and
ha
v
e
s
pu
rr
ed
p
i
onee
ri
n
g
w
o
r
k
s
t
ha
t
s
ou
g
h
t
t
o
ha
r
ne
ss
t
he
po
w
e
r
o
f
neu
r
a
l
ne
t
w
o
r
k
s
f
o
r
N
L
P
t
a
s
k
s. R
e
c
en
t
l
y
,
t
he
i
n
t
r
odu
c
t
i
on
o
f
t
he
a
tt
en
t
i
on
m
e
c
han
ism
and
t
r
an
s
f
o
rm
e
r
ne
t
w
o
r
k
s
ha
s
enab
l
ed
a
sig
n
i
f-
ic
an
t
s
h
i
f
t
a
w
a
y
f
r
o
m r
e
c
u
rr
en
t
and
c
on
v
o
l
u
t
i
ona
l m
e
t
hod
s,
t
r
ad
i
t
i
ona
ll
y
u
s
ed
i
n
deep
-
l
ea
r
n
i
n
g
t
a
s
k
s
[
8
]
. T
h
is
ha
s l
ed
t
o
t
he
de
v
e
l
op
m
en
t
o
f
t
r
an
s
f
o
rm
e
r
-
ba
s
ed
m
ode
ls, s
u
c
h
a
s BERT
[
9
]
, w
h
ic
h
o
wi
n
g
t
o
i
t
s c
on
t
e
x
t
ua
l c
o
m
p
r
ehen
si
on
o
f
l
an
g
ua
g
e
ha
s
been
u
s
ed
e
x
t
en
si
v
e
l
y
f
o
r s
en
t
im
en
t
ana
l
y
sis.
H
o
w
e
v
e
r,
t
he
pe
r
f
o
rm
an
c
e
o
f
BERT i
n
t
he
f
i
nan
ci
a
l
do
m
a
i
n
ha
s
en
c
oun
t
e
r
ed
limi
t
a
t
i
on
s,
p
rim
a
ril
y
be
c
au
s
e
i
t
is
no
t
s
pe
ci
f
ic
a
ll
y
t
r
a
i
ned
on
f
i
nan
ci
a
l
da
t
a
s
e
t
s. M
o
r
eo
v
e
r, i
t
s r
equ
ir
e
m
en
t
f
o
r s
ub
s
t
an
t
i
a
l
a
m
oun
t
s
o
f
da
t
a
f
o
r
f
i
ne
-
t
un
i
n
g
pu
r
po
s
e
s
po
s
e
s
a
c
on
si
de
r
ab
l
e
c
ha
ll
en
g
e
f
o
r
f
i
nan
ci
a
l
app
lic
a
t
i
on
s, w
he
r
e
s
u
c
h
da
t
a
m
a
y
no
t
be
r
ead
il
y
a
v
a
il
ab
l
e
.
M
o
r
e
r
e
c
en
t
l
y
, Fi
n
BERT
[
10
]
,
a
v
e
rsi
on
o
f
BERT w
h
ic
h
is
f
i
ne
-
t
uned
on
f
i
nan
ci
a
l
t
e
x
t
,
ha
s s
ho
w
n
p
r
o
misi
n
g r
e
s
u
l
t
s
f
o
r
t
he
t
a
s
k
o
f
f
i
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis.
H
o
w
e
v
e
r, Fi
n
BERT
s
t
ill s
u
ff
e
rs
f
r
o
m limi
t
a
t
i
on
s s
u
c
h
a
s i
n
s
en
si
t
i
v
i
t
y
t
o
nu
m
e
ric
a
l
v
a
l
ue
s, w
h
il
e
due
t
o
i
t
s r
e
l
a
t
i
v
e
l
y
sm
a
ll si
z
e
(
110 milli
on
pa
-
r
a
m
e
t
e
rs
)
i
t
s cl
a
ssi
f
ic
a
t
i
on
a
cc
u
r
a
c
y
de
t
e
ri
o
r
a
t
e
s wi
t
h
s
en
t
en
c
e
c
o
m
p
l
e
x
i
t
y
[
11
]
. T
he
Fi
n
GP
T
[
12
]
,
[
13
]
and
I
n
s
t
r
u
c
t
-
Fi
n
GP
T
[
14
]
a
im
t
o
enhan
c
e
t
he
ir
e
x
p
r
e
ssi
v
e
po
w
e
r
b
y
u
si
n
g
t
he
Ll
a
m
a
7B
a
s
t
he
ir
ba
s
e
m
ode
l.
H
o
w
e
v
e
r, Fi
n
GP
T is
no
t
op
t
imi
z
ed
f
o
r
t
he
t
a
s
k
o
f
f
i
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis w
h
ils
t
I
n
s
t
r
u
c
t
-
Fi
n
GP
T
on
l
y
cl
a
ssi
f
i
e
s
t
he
s
en
t
im
en
t
v
a
l
en
c
e
bu
t
is
no
t
c
apab
l
e
o
f
quan
t
i
fy
i
n
g
t
he
s
t
r
en
g
t
h
o
f
a
s
en
t
im
en
t
cl
a
ss.
T
o
t
he
be
s
t
o
f
ou
r
k
no
wl
ed
g
e
, Bl
oo
m
be
rg
GP
T
[
15
]
is
t
he
on
l
y
p
r
e
-
t
r
a
i
ned
f
i
nan
c
e
-
s
pe
ci
f
ic LLM,
a
s Bl
oo
m
be
rg w
a
s
ab
l
e
t
o
t
r
a
i
n
t
he
m
ode
l
u
si
n
g
da
t
a
c
o
ll
e
c
t
ed
o
v
e
r
a
s
pan
o
f
40
y
ea
rs.
D
e
s
p
i
t
e
t
he
im
p
r
e
ssi
v
e
pe
r
f
o
rm
an
c
e
o
f
t
he
m
ode
l
on
f
i
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis,
t
he
r
e
s
ou
rc
e
s r
equ
ir
ed
t
o
t
r
a
i
n
s
u
c
h
a
m
ode
l
a
r
e
s
ub
s
t
an
t
i
a
l
(
1.3M
GPU
hou
rs
a
t
a
c
o
s
t
o
f
$5M
)
w
h
ils
t
m
u
c
h
o
f
t
he
t
r
a
i
n
i
n
g
da
t
a
is c
onf
i
den
t
i
a
l
and
no
t
pub
licl
y
a
v
a
il
ab
l
e
. T
h
is is
d
i
ff
e
r
en
t
f
r
o
m
ou
r
p
r
opo
s
ed
m
e
t
hodo
l
o
g
y
, w
h
ic
h
f
o
c
u
s
e
s
on
a
c
h
i
e
v
i
n
g
a
h
ig
h
cl
a
ssi
f
ic
a
t
i
on
a
cc
u
r
a
c
y
w
h
ils
t
mi
n
imi
z
i
n
g
t
he
t
r
a
i
n
i
n
g c
o
r
pu
s
and
c
o
m
pu
t
a
-
t
i
ona
l r
e
s
ou
rc
e
s. T
h
is is
a
c
h
i
e
v
ed
b
y
f
i
ne
-
t
un
i
n
g
a
p
r
e
-
t
r
a
i
ned
g
ene
r
a
l
-
pu
r
po
s
e
LLM
on
a
sm
a
ll
e
r
-
sc
a
l
e
f
i
nan
ci
a
l
da
t
a
c
o
r
pu
s.
III. M
ET
H
O
D
O
L
OG
Y
O
u
r w
o
r
k
a
ims
t
o
e
m
ba
r
k
upon
t
he
imm
en
s
e
e
x
p
r
e
ssi
v
e
po
w
e
r
and
c
on
t
e
x
t
ua
l
unde
rs
t
and
i
n
g
o
f
g
ene
r
a
l
-
pu
r
po
s
e
LLMs i
n
o
r
de
r
t
o
m
a
k
e
t
he
m
f
i
nan
c
e
-
s
pe
ci
f
ic. T
h
is is
a
c
h
i
e
v
ed
b
y
f
i
ne
-
t
un
i
n
g
t
he
s
t
a
t
e
-
o
f-
t
he
-
a
r
t
(
S
O
TA
)
Ll
a
m
a
2 7B m
ode
l
on
a
s
pe
ci
f
ic c
o
r
pu
s
o
f
f
i
nan
ci
a
l
da
t
a
. T
he
e
ff
e
c
t
i
v
ene
ss
o
f
ou
r
m
ode
l is
de
m
on
s
t
r
a
t
ed
on
f
i
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis
t
h
r
ou
g
h
a
ne
w s
e
t
o
f
ben
c
h
m
a
r
k
s
t
ha
t
a
lig
n
cl
o
s
e
l
y
wi
t
h
end
po
r
t
f
o
li
o
c
on
s
t
r
u
c
t
i
on
-
t
he
u
l
t
im
a
t
e
g
oa
l
o
f
f
i
nan
ci
a
l
ana
l
y
sis.
A. Fi
ne
-
t
un
i
n
g
t
he
Lla
m
a 2
mode
l
E
v
en
t
hou
g
h
p
r
e
-
t
r
a
i
ned
LLMs
po
s
e
a
r
an
g
e
o
f
c
apab
ili
t
i
e
s
s
u
c
h
a
s r
ea
s
on
i
n
g,
t
r
an
sl
a
t
i
on
, s
u
mm
a
risi
n
g
and
t
e
x
t
g
ene
r
a
-
t
i
on
,
t
he
y
o
f
t
en
s
t
r
u
ggl
e
w
hen
app
li
ed
t
o
a
s
pe
ci
f
ic
t
a
s
k
o
f
i
n
t
e
r
e
s
t
, s
u
c
h
a
s s
en
t
im
en
t
ana
l
y
sis. T
h
is limi
t
a
t
i
on
be
c
o
m
e
s
e
v
en
m
o
r
e
cri
t
ic
a
l i
n
t
he
f
i
nan
c
e
do
m
a
i
n
, w
he
r
e
t
he
nuan
c
ed
l
an
g
ua
g
e
, m
ed
i
a
h
y
pe
and
e
x
t
en
si
v
e
l
en
g
t
h
o
f
f
i
nan
ci
a
l
ne
ws
a
r
t
icl
e
s
po
s
e
sig
n
i
f
ic
an
t
add
i
t
i
ona
l c
ha
ll
en
g
e
s.
T
o
t
a
c
k
l
e
t
he
s
e
c
ha
ll
en
g
e
s,
ou
r w
o
r
k
r
e
v
isi
t
s
t
he
f
irs
t
p
ri
n
-
ci
p
l
e
s
o
f
LLMs i
n
o
r
de
r
t
o
a
lig
n
t
he
m
t
o
t
he
t
a
s
k
o
f
f
i
nan
ci
a
l
s
en
t
im
en
t
ana
l
y
sis. T
h
is is
a
c
h
i
e
v
ed
b
y
u
si
n
g
f
ou
r l
abe
ll
ed
f
i
nan
ci
a
l
t
e
x
t
da
t
a
s
e
t
s
a
s
t
r
a
i
n
i
n
g
da
t
a
t
o
f
i
ne
-
t
une
t
he
Ll
a
m
a
2
m
ode
l. S
u
c
h
t
r
a
i
n
i
n
g
on
f
i
nan
ci
a
l
da
t
a
,
equ
i
pped
t
he
m
ode
l wi
t
h
t
he
ab
ili
t
y
t
o
unde
rs
t
and
t
he
li
n
g
u
is
t
ic
nuan
c
e
s
p
r
e
s
en
t
i
n
t
he
f
i
nan
ci
a
l
do
m
a
i
n
. F
u
r
t
he
rm
o
r
e
,
a
S
o
f
t
M
a
x
cl
a
ssi
f
ic
a
t
i
on
l
a
y
e
r
w
a
s
added
a
t
t
he
ou
t
pu
t
o
f
t
he
f
ounda
t
i
ona
l m
ode
l,
a
ll
o
wi
n
g
t
he
p
r
opo
s
ed
f
i
ne
-
t
uned
m
ode
l
t
o
p
r
odu
c
e
S
o
f
t
M
a
x
ou
t
pu
t
s
f
o
r
t
h
r
ee
l
abe
ls:
po
si
t
i
v
e
,
ne
g
a
t
i
v
e
o
r
neu
t
r
a
l. T
h
is m
ade
i
t
po
ssi
b
l
e
t
o
a
l
t
e
r
t
he
p
rim
a
r
y
f
un
c
t
i
on
o
f
t
he
m
ode
l
f
r
o
m
t
e
x
t
g
ene
r
a
t
i
on
t
o
s
en
t
im
en
t
cl
a
ssi
f
ic
a
t
i
on
.
1
)
Tr
ai
n
i
n
g
d
a
t
as
e
t
s:
O
u
r
t
r
a
i
n
i
n
g
da
t
a
w
a
s
a
c
o
m
b
i
na
t
i
on
o
f
f
ou
r l
abe
ll
ed
pub
licl
y
a
v
a
il
ab
l
e
f
i
nan
ci
a
l
ne
ws
da
t
a
s
e
t
s,
r
e
s
u
l
t
i
n
g i
n
a
c
o
m
p
r
ehen
si
v
e
c
o
ll
e
c
t
i
on
o
f
34,180 l
abe
ll
ed
s
a
m
p
l
e
s. E
a
c
h
s
a
m
p
l
e
w
a
s
anno
t
a
t
ed
wi
t
h
one
o
f
t
he
t
h
r
ee
l
abe
ls:
po
si
t
i
v
e
,
ne
g
a
t
i
v
e
,
and
neu
t
r
a
l.
2
)
M
ode
l
Tr
ai
n
i
n
g:
O
u
r Fi
n
Ll
a
m
a
m
ode
l w
a
s
f
irs
t
i
n
i
-
t
i
a
lis
ed
wi
t
h
t
he
Ll
a
m
a
2 7B m
ode
l,
f
o
ll
o
w
ed
b
y
f
i
ne
-
t
un
i
n
g
o
v
e
r 5
epo
c
h
s. T
he
t
r
a
i
n
i
n
g
p
r
o
c
e
ss
u
t
ilis
ed
t
he
A
da
mW
op
t
imi
z
e
r
[
16
]
,
a
s i
t
e
ff
e
c
t
i
v
e
l
y
de
c
oup
l
e
s
t
he
w
e
ig
h
t
de
c
a
y
f
r
o
m
t
he
op
t
imi
z
a
t
i
on
s
t
ep
s, l
ead
i
n
g
t
o
m
o
r
e
e
ff
e
c
t
i
v
e
t
r
a
i
n
i
n
g.
T
he
i
n
i
t
i
a
l l
ea
r
n
i
n
g r
a
t
e
w
a
s
de
li
be
r
a
t
e
l
y
k
ep
t
sm
a
ll
a
s
t
he
Ll
a
m
a
2 7B m
ode
l is
a
lr
ead
y
p
r
e
-
t
r
a
i
ned
on
a
l
a
rg
e
c
o
r
pu
s
o
f
da
t
a
, w
h
ils
t
t
he
w
a
rm
-
up
r
a
t
i
o
and
w
e
ig
h
t
de
c
a
y
s
e
r
v
ed
a
s
k
e
y
r
e
g
u
l
a
ris
a
t
i
on
t
e
c
hn
i
que
s
t
o
p
r
e
v
en
t
o
v
e
r
f
i
tt
i
n
g,
a
cr
u
ci
a
l
a
s
pe
c
t
gi
v
en
t
he
limi
t
ed
si
z
e
o
f
ou
r
f
i
ne
-
t
un
i
n
g
da
t
a
s
e
t
.
2
Fig. 1:
Ov
e
r
v
i
e
w
o
f
s
en
t
im
en
t
ana
l
y
sis m
e
t
hod
s.
M
o
r
eo
v
e
r,
t
he
L
o
RA im
p
l
e
m
en
t
a
t
i
on
w
a
s
e
m
p
l
o
y
ed
i
n
t
he
f
i
ne
-
t
un
i
n
g
p
r
o
c
e
ss i
n
o
r
de
r
t
o
mi
n
imi
z
e
t
he
nu
m
be
r
o
f
t
r
a
i
nab
l
e
pa
r
a
m
e
t
e
rs w
h
ils
t
a
c
h
i
e
v
i
n
g
h
ig
h
and
r
obu
s
t
end
pe
r
f
o
rm
an
c
e
. T
h
r
ou
g
h
t
he
L
o
RA im
p
l
e
m
en
t
a
t
i
on
,
t
he
nu
m
be
r
o
f
t
r
a
i
nab
l
e
pa
r
a
m
e
t
e
rs w
a
s s
e
t
t
o
4.2M,
a
m
oun
t
i
n
g
t
o
j
u
s
t
0.0638
%
o
f
t
he
t
o
t
a
l
nu
m
be
r
o
f
pa
r
a
m
e
t
e
rs i
n
t
he
Ll
a
m
a
2 7B
m
ode
l. T
h
is m
ade
i
t
po
ssi
b
l
e
f
o
r
ou
r
f
i
ne
-
t
un
i
n
g
p
r
o
c
e
ss
t
o
be
im
p
l
e
m
en
t
ed
on
a
singl
e
A100
(
40
G
B
)
GPU
,
t
hu
s
a
v
o
i
d
i
n
g
t
he
need
f
o
r
e
x
c
e
ssi
v
e
c
o
m
pu
t
a
t
i
ona
l r
e
s
ou
rc
e
s.
B
.
P
r
opo
s
ed
F
r
a
mewo
r
k
Wi
t
h
t
he
p
r
opo
s
ed
f
i
ne
-
t
uned
Ll
a
m
a
2 m
ode
l i
n
p
l
a
c
e
, w
e
f
o
ll
o
w
ed
t
he
f
r
a
m
e
w
o
r
k
s
ho
w
n
i
n
Fig
u
r
e
2.
O
u
r
a
im w
a
s
t
o
a
ss
e
ss
t
he
pe
r
f
o
rm
an
c
e
o
f
ou
r Fi
n
Lla
m
a m
ode
l
a
g
a
i
n
s
t
o
t
he
r
e
s
t
ab
lis
hed
s
en
t
im
en
t
ana
l
y
sis m
e
t
hod
s,
u
si
n
g
f
i
nan
c
e
-
s
pe
ci
f
ic
r
ea
l
-
w
o
rl
d
m
e
t
rics.
Fig. 2: Fr
a
m
e
w
o
r
k
f
o
r s
en
t
im
en
t
ana
l
y
sis.
D
ata C
o
ll
e
cti
o
n and
P
r
o
c
e
ssing. B
o
t
h
t
e
x
t
ua
l
and
m
a
r
k
e
t
da
t
a
w
e
r
e
c
o
ll
e
c
t
ed
i
n
o
r
de
r
t
o
c
on
s
t
r
u
c
t
app
r
op
ri
a
t
e
l
on
g
-
s
ho
r
t
(
L/S
)
po
r
t
f
o
li
o
s. R
e
g
a
r
d
i
n
g
t
he
t
e
x
t
ua
l
da
t
a
, 204,017
a
r
t
icl
e
s
da
t
i
n
g
be
t
w
een
2015
t
o
2021 w
e
r
e
c
o
ll
e
c
t
ed
f
r
o
m
on
li
ne
s
ou
rc
e
s. T
he
s
e
s
ou
rc
e
s w
e
r
e
s
e
l
e
c
t
ed
due
t
o
t
he
ir r
e
li
ab
ili
t
y
,
r
epu
t
a
t
i
on
, l
a
c
k
o
f
b
i
a
s
and
f
o
c
u
s
on
m
a
j
o
r c
o
r
po
r
a
t
i
on
s.
Fi
nan
ci
a
l m
a
r
k
e
t
da
t
a
w
e
r
e
c
o
ll
e
c
t
ed
f
o
r
t
he
s
a
m
e
t
im
e
pe
ri
od
f
r
o
m Y
ahoo
Fi
nan
c
e
. T
he
s
e
c
o
ll
e
c
t
ed
m
a
r
k
e
t
da
t
a
c
on
t
a
i
ned
da
il
y
s
t
o
c
k
r
e
t
u
r
n
s
f
o
r
t
he
500 c
o
m
pan
i
e
s i
n
ou
r I
n
v
e
s
t
ab
l
e
U
n
i
v
e
rs
e
(
S
&P
500
)
, r
e
s
u
l
t
i
n
g i
n
1,672
da
y
s
o
f
s
t
o
c
k
r
e
t
u
r
n
s
da
t
a
f
o
r
ea
c
h
c
o
m
pan
y
.
D
a
t
a
p
r
o
c
e
ssi
n
g i
n
t
he
f
o
rm
o
f
N
a
m
ed
E
n
t
i
t
y
R
e
c
o
g
n
i
t
i
on
(N
ER
)
and
t
e
x
t
p
r
e
-
p
r
o
c
e
ssi
n
g w
a
s
t
hen
app
li
ed
t
o
t
he
t
e
x
t
ua
l
da
t
a
,
t
o
r
e
m
o
v
e
irr
e
l
e
v
an
t
a
r
t
icl
e
s
and
en
s
u
r
e
t
he
c
o
m
pa
t
i
b
ili
t
y
o
f
t
he
a
r
t
icl
e
s wi
t
h
ou
r s
en
t
im
en
t
m
e
t
hdod
s.
S
e
ntim
e
nt Anal
y
sis. I
n
t
o
t
a
l,
f
i
v
e
s
en
t
im
en
t
ana
l
y
sis m
e
t
h
-
od
s w
e
r
e
app
li
ed
. F
o
r
t
he
l
e
x
ic
on
-
ba
s
ed
app
r
oa
c
he
s
(
s
ee
A
p
-
pend
i
x
VI
-
A1
)
, LM
D
[
17
]
and
H
IV
-
4
[
18
]
w
e
r
e
im
p
l
e
m
en
t
ed
u
si
n
g
t
he
p
y
s
en
t
im
en
t
2
Py
t
hon
li
b
r
a
r
y
, w
h
il
e
VA
D
ER
[
19
]
w
a
s im
p
l
e
m
en
t
ed
u
si
n
g
t
he
N
LTK li
b
r
a
r
y
. R
e
g
a
r
d
i
n
g
t
he
deep
l
ea
r
n
i
n
g m
e
t
hod
s
(
s
ee
A
ppend
i
x
VI
-
B
)
,
bo
t
h
t
he
Fi
n
BERT
m
ode
l
and
ou
r Fi
n
Ll
a
m
a
m
ode
l w
e
r
e
ob
t
a
i
ned
t
h
r
ou
g
h
H
u
g
-
gi
n
gF
a
c
e
,
and
w
e
r
e
u
t
ilis
ed
v
i
a
t
he
Tr
an
s
f
o
rm
e
rs li
b
r
a
r
y
.
T
he
c
on
si
de
r
ed
m
e
t
hod
s w
e
r
e
e
v
a
l
ua
t
ed
on
e
v
e
r
y
a
r
t
icl
e
wi
t
h
i
n
ea
c
h
c
o
r
pu
s
f
o
r
a
gi
v
en
c
o
m
pan
y
. I
n
c
a
s
e
s w
he
r
e
m
u
l
t
i
p
l
e
a
r
t
icl
e
s w
e
r
e
pub
lis
hed
on
t
he
s
a
m
e
da
y
f
o
r
a
gi
v
en
c
o
m
pan
y
,
t
he
a
v
e
r
a
g
e
s
en
t
im
en
t
w
a
s c
a
lc
u
l
a
t
ed
f
o
r
t
ha
t
da
y
, i
n
t
he
f
o
rm
N
N
t
S
t
=
1
X
S
it
(
1
)
t
i=
1
H
e
r
e
, S
t
r
ep
r
e
s
en
t
s
t
he
a
v
e
r
a
g
e
s
en
t
im
en
t
f
o
r
t
he
t
-
t
h
da
y
, N
t
deno
t
e
s
t
he
nu
m
be
r
o
f
ne
ws
a
r
t
icl
e
s
pub
lis
hed
on
t
ha
t
s
a
m
e
t
-
t
h
da
y
f
o
r
a
gi
v
en
c
o
m
pan
y
, w
h
il
e
S
it
de
sig
na
t
e
s
t
he
s
en
t
im
en
t
s
t
r
en
g
t
h
o
f
t
he
i
-
t
h
ne
ws
a
r
t
icl
e
on
a
pa
r
t
ic
u
l
a
r
t
-
t
h
da
y
. T
he
da
il
y
s
en
t
im
en
t
ou
t
pu
t
s
f
o
r
ea
c
h
c
o
m
pan
y
w
e
r
e
m
e
rg
ed
t
o
a
rri
v
e
a
t
t
he
f
i
na
l s
en
t
im
en
t
da
t
a
t
ha
t
w
e
r
e
u
t
ilis
ed
a
s
a
pa
r
a
m
e
t
e
r i
n
t
he
po
r
t
f
o
li
o
c
on
s
t
r
u
c
t
i
on
s
t
a
g
e
.
3
P
o
rt
f
o
li
o
C
o
nstructi
o
n.
O
n
c
e
t
he
s
en
t
im
en
t
f
o
r
ea
c
h
m
e
t
hod
w
a
s
def
i
ned
f
o
r
e
v
e
r
y
c
o
m
pan
y
,
t
he
l
on
g
-
s
ho
r
t
po
r
t
-
f
o
li
o
w
a
s c
on
s
t
r
u
c
t
ed
. W
e
u
s
ed
t
he
s
en
t
im
en
t
a
s
a
pa
r
a
m
e
t
e
r
t
o
de
t
e
rmi
ne
w
h
ic
h
c
o
m
pan
i
e
s s
hou
l
d
be
i
n
a
l
on
g
o
r s
ho
r
t
po
si
t
i
on
,
a
imi
n
g
t
o
m
a
x
imis
e
r
e
t
u
r
n
s
f
r
o
m
bo
t
h
po
si
t
i
on
s.
T
he
l
on
g
-
s
ho
r
t
po
r
t
f
o
li
o
w
a
s c
on
s
t
r
u
c
t
ed
u
si
n
g
t
he
f
o
ll
o
wi
n
g
p
r
o
c
edu
r
e
:
•
D
e
f
i
ne
t
he
I
n
v
e
s
t
a
b
l
e
U
n
iv
e
r
s
e
: E
v
en
t
hou
g
h
t
he
S
&P
500 c
o
m
p
ris
e
s 500 c
o
m
pan
i
e
s,
t
he
f
i
nan
ci
a
l
t
e
x
t
ua
l
da
t
a
c
o
ll
e
c
t
ed
d
i
d
no
t
c
on
t
a
i
n
a
r
t
icl
e
s
a
ss
o
ci
a
t
ed
t
o
s
o
m
e
o
f
t
he
c
o
m
pan
i
e
s
f
o
r
t
he
t
e
s
t
pe
ri
od
o
f
F
eb
r
ua
r
y
2015
t
o
J
une
2021. C
on
s
equen
t
l
y
, 417 c
o
m
pan
i
e
s w
e
r
e
c
on
si
de
r
ed
.
•
D
e
f
i
ne
t
he
l
on
g a
nd
s
ho
r
t
po
si
t
i
on
: T
he
s
en
t
im
en
t
sig
-
na
l
ob
t
a
i
ned
f
r
o
m
ea
c
h
o
f
t
he
f
i
v
e
m
e
t
hod
s w
a
s
u
s
ed
t
o
c
on
s
t
r
u
c
t
f
i
v
e
d
is
t
i
n
c
t
po
r
t
f
o
li
o
s. F
o
r
ea
c
h
m
e
t
hod
,
c
o
m
pan
i
e
s w
e
r
e
r
an
k
ed
da
il
y
a
cc
o
r
d
i
n
g
t
o
t
he
ir s
en
t
i
-
m
en
t
. C
o
m
pan
i
e
s
t
ha
t
d
i
d
no
t
ha
v
e
s
en
t
im
en
t
da
t
a
on
a
pa
r
t
ic
u
l
a
r
da
y
w
e
r
e
o
mi
tt
ed
f
r
o
m
t
he
r
an
k
i
n
g. As
t
he
da
il
y
s
en
t
im
en
t
sc
o
r
e
f
o
r
ea
c
h
c
o
m
pan
y
r
an
g
e
s
be
t
w
een
-
1
and
1,
t
ho
s
e
wi
t
h
t
he
h
ig
he
s
t
po
si
t
i
v
e
s
en
t
im
en
t
w
e
r
e
p
l
a
c
ed
i
n
l
on
g
po
si
t
i
on
s, w
h
ils
t
t
ho
s
e
wi
t
h
t
he
s
t
r
on
g
e
s
t
ne
g
a
t
i
v
e
s
en
t
im
en
t
w
e
r
e
p
l
a
c
ed
i
n
s
ho
r
t
po
si
t
i
on
s.
•
All
o
ca
t
i
on
: A
n
equa
ll
y-
w
e
ig
h
t
ed
po
r
t
f
o
li
o
s
t
r
a
t
e
g
y
w
a
s
c
on
si
de
r
ed
i
n
ou
r
po
r
t
f
o
li
o
c
on
s
t
r
u
c
t
i
on
a
s
t
h
is is
t
he
s
t
r
a
t
-
e
g
y
m
o
s
t
l
y
u
t
ilis
ed
b
y
hed
g
e
f
und
s
[
20
]
. T
he
pe
rc
en
t
a
g
e
o
f
c
o
m
pan
i
e
s i
n
a
l
on
g
and
s
ho
r
t
po
si
t
i
on
w
a
s
f
i
x
ed
a
t
35
%
.
C
on
s
equen
t
l
y
,
t
he
t
op
35
%
o
f
c
o
m
pan
i
e
s i
n
t
e
rms
o
f
pe
r
f
o
rm
an
c
e
w
e
r
e
a
ll
o
c
a
t
ed
t
o
l
on
g
po
si
t
i
on
s, w
h
il
e
t
he
bo
tt
o
m 35
%
w
e
r
e
a
ll
o
c
a
t
ed
t
o
s
ho
r
t
po
si
t
i
on
s.
•
D
e
t
e
r
m
i
ne
d
aily
r
e
t
u
r
n
s: T
he
da
il
y
r
e
t
u
r
n
f
o
r
ea
c
h
c
o
m
-
pan
y
t
ha
t
w
a
s
he
l
d
i
n
a
l
on
g
o
r s
ho
r
t
po
si
t
i
on
w
a
s
ob
t
a
i
ned
b
y
t
he
m
a
r
k
e
t
da
t
a
on
t
ha
t
pa
r
t
ic
u
l
a
r
da
y
. T
he
t
o
t
a
l
da
il
y
r
e
t
u
r
n
o
f
c
o
m
pan
i
e
s
t
ha
t
w
e
r
e
he
l
d
i
n
a
l
on
g
po
si
t
i
on
w
a
s
def
i
ned
a
s
X
N
D
a
il
y
L
o
n
g R
e
tu
r
n
r
l
o
n
g
=
N
long
r
l
o
n
g
(i)
(
2
)
i=
1
l
o
n
g
Simil
a
rl
y
,
t
he
t
o
t
a
l
da
il
y
r
e
t
u
r
n
o
f
c
o
m
pan
i
e
s
t
ha
t
w
e
r
e
he
l
d
i
n
a
s
ho
r
t
po
si
t
i
on
w
a
s
def
i
ned
a
s
X
N
D
a
il
y
S
h
o
r
t
R
e
tu
r
n
r
sh
o
rt
=
N
short
r
sh
o
rt
(i)
(
3
)
i=
1
sh
o
rt
F
o
r
ea
c
h
pa
r
t
ic
u
l
a
r
da
y
,
t
he
nu
m
be
r
o
f
c
o
m
pan
i
e
s
t
ha
t
w
e
r
e
he
l
d
i
n
e
i
t
he
r
a
l
on
g
po
si
t
i
on
(
N
l
o
n
g
)
o
r
a
s
ho
r
t
po
si
t
i
on
(
N
sh
o
rt
)
w
e
r
e
equa
l. C
on
s
equen
t
l
y
,
t
he
t
o
t
a
l
po
r
t
f
o
li
o
r
e
t
u
r
n
on
a
pa
r
t
ic
u
l
a
r
da
y
w
a
s
t
he
d
i
ff
e
r
en
c
e
be
t
w
een
t
he
da
il
y
l
on
g r
e
t
u
r
n
and
da
il
y
s
ho
r
t
r
e
t
u
r
n
,
and
is
gi
v
en
b
y
D
a
il
y
R
e
tu
r
n
r
d
aily
(i) = r
l
o
n
g
(i) − r
sh
o
rt
(i)
(
4
)
P
o
rt
f
o
li
o
E
v
aluati
o
n. T
he
pe
r
f
o
rm
an
c
e
o
f
t
he
po
r
t
f
o
li
o
c
on
-
s
t
r
u
c
t
ed
u
si
n
g
ou
r
f
i
ne
-
t
uned
m
ode
l w
a
s
a
ss
e
ss
ed
a
g
a
i
n
s
t
t
he
po
r
t
f
o
li
o
s c
on
s
t
r
u
c
t
ed
u
si
n
g
o
t
he
r S
O
TA s
en
t
im
en
t
m
e
t
hod
s.
T
o
t
h
is
end
,
t
he
e
m
p
l
o
y
ed
r
ea
l
-
w
o
rl
d
f
i
nan
ci
a
l m
e
t
rics w
e
r
e
:
c
u
m
u
l
a
t
i
v
e
r
e
t
u
r
n
s,
annua
li
z
ed
r
e
t
u
r
n
,
annua
li
z
ed
v
o
l
a
t
ili
t
y
,
and
t
he
S
ha
r
pe
r
a
t
i
o
[
21
]
. T
he
s
e
m
e
t
rics
a
r
e
def
i
ned
a
s
X
N
C
umu
l
at
i
v
e
R
e
tu
r
n
s
r
cum
=r
d
aily
(i)
(
5
)
i=
1
X
N
A
nnua
li
z
e
d R
e
tu
r
n
R
p
=
N
r
l
o
g
(i)
× 252
(
6
)
i=
1
P
N
N − 1
s
A
nn.
V
o
l
at
ili
ty
σ
p
=
i=
1
(r
l
o
g
(i)
−
r&hibar;)
2
×
√
252
(
7
)
σ
S
ha
rp
e
R
at
i
o
S
a
= R
p
−
R
f
(
8
)
p
w
he
r
e
N is
t
he
t
o
t
a
l
nu
m
be
r
o
f
i
n
v
e
s
t
i
n
g
da
y
s,
t
o
t
a
li
n
g 1,672,
r
l
o
g
(i) is
t
he
l
o
g
a
ri
t
h
mic
da
il
y
r
e
t
u
r
n
, r&hibar; is
t
he
a
v
e
r
a
g
e
da
il
y
l
o
g
a
ri
t
h
mic r
e
t
u
r
n
, R
f
is
t
he
annua
li
z
ed
ris
k-f
r
ee
r
a
t
e
o
f
r
e
t
u
r
n
and
252 is
t
he
nu
m
be
r
o
f
bu
si
ne
ss
da
y
s i
n
a
y
ea
r. T
he
ris
k-f
r
ee
r
e
t
u
r
n
, R
f
,
t
y
p
ic
a
ll
y
r
ep
r
e
s
en
t
s
t
he
y
i
e
l
d
o
f
t
he
10
-
Y
ea
r
Tr
ea
s
u
r
y
N
o
t
e
;
ho
w
e
v
e
r,
due
t
o
i
t
s
p
r
o
l
on
g
ed
l
o
w
y
i
e
l
d
[
22
]
du
ri
n
g
t
he
ana
l
y
s
ed
pe
ri
od
,
a
0
%
r
a
t
e
is c
o
mm
on
l
y
u
s
ed
and
is
adop
t
ed
i
n
ou
r
ana
l
y
sis.
IV. E
X
P
ERIME
N
TAL
R
ES
U
LTS
T
he
pe
r
f
o
rm
an
c
e
o
f
t
he
f
i
v
e
po
r
t
f
o
li
o
s w
h
ic
h
w
e
r
e
c
on
-
s
t
r
u
c
t
ed
a
s
de
scri
bed
i
n
S
e
c
t
i
on
III
a
r
e
s
ho
w
n
i
n
Fig
u
r
e
3.
O
b
s
e
r
v
e
t
ha
t
t
he
deep
l
ea
r
n
i
n
g
app
r
oa
c
he
s
ou
t
pe
r
f
o
rm
ed
t
he
l
e
x
ic
on
-
ba
s
ed
app
r
oa
c
he
s i
n
t
e
rms
o
f
c
u
m
u
l
a
t
i
v
e
r
e
t
u
r
n
s,
pa
r
-
t
ic
u
l
a
rl
y
t
ho
s
e
r
e
l
y
i
n
g
on
g
ene
r
a
l
-
pu
r
po
s
e
d
ic
t
i
ona
ri
e
s
(
H
IV
-
4
and
V
A
D
ER
)
. T
h
is w
a
s
t
o
be
e
x
pe
c
t
ed
, gi
v
en
t
ha
t
l
e
x
ic
on
-
ba
s
ed
app
r
oa
c
he
s
o
f
t
en
f
a
il
t
o
c
ap
t
u
r
e
t
he
c
on
t
e
x
t
ua
l m
ean
i
n
g
o
f
s
en
t
en
c
e
s, w
h
ils
t
t
he
nuan
c
ed
na
t
u
r
e
o
f
f
i
nan
ci
a
l
t
e
x
t
sig
n
i
f-
ic
an
t
l
y
r
edu
c
e
s
t
he
a
cc
u
r
a
c
y
o
f
g
ene
r
a
l
-
pu
r
po
s
e
d
ic
t
i
ona
ri
e
s.
F
u
r
t
he
rm
o
r
e
,
a
s
ob
s
e
r
v
ed
i
n
t
he
bo
tt
o
m
-
l
e
f
t
pane
l
o
f
Fig
u
r
e
3,
a
ll m
e
t
hod
s
e
x
h
i
b
i
t
ed
t
he
ir
be
s
t
pe
r
f
o
rm
an
c
e
du
ri
n
g
t
u
r
bu
l
en
t
and
h
ig
h
-v
o
l
a
t
ili
t
y
e
c
ono
mic
pe
ri
od
s, s
u
c
h
a
s
t
he
f
irs
t
qua
r
t
e
r
o
f
2020. T
he
r
e
s
u
l
t
s
ob
s
e
r
v
ed
i
n
T
ab
l
e
II s
u
gg
e
s
t
t
ha
t
t
he
35
%
l
on
g
-
s
ho
r
t
po
r
t
f
o
li
o
, c
on
s
t
r
u
c
t
ed
u
si
n
g
ou
r
f
i
ne
-
t
uned
Ll
a
m
a
-
2 m
ode
l,
w
a
s
t
he
m
o
s
t
s
u
cc
e
ss
f
u
l. T
h
is is
a
tt
ri
bu
t
ed
t
o
i
t
s
ab
ili
t
y
t
o
a
c
h
i
e
v
e
sig
n
i
f
ic
an
t
l
y
h
ig
he
r c
u
m
u
l
a
t
i
v
e
r
e
t
u
r
n
s c
o
m
pa
r
ed
t
o
a
ll
o
t
he
r c
on
si
de
r
ed
m
e
t
hod
s,
and
m
o
s
t
im
po
r
t
an
t
l
y
Fi
n
BER
T
, w
h
il
e
a
tt
a
i
n
i
n
g
a
h
ig
he
r S
ha
r
pe
r
a
t
i
o
and
e
x
h
i
b
i
t
i
n
g l
o
w
e
r
v
o
l
a
t
ili
t
y
.
Ov
e
r
a
ll,
ou
r Fi
n
Ll
a
m
a
m
ode
l s
u
cc
e
ss
f
u
ll
y
g
ene
r
a
t
ed
sig
-
n
i
f
ic
an
t
l
y
h
ig
he
r r
e
t
u
r
n
s
f
o
r i
n
v
e
s
t
o
rs, w
h
ils
t
sim
u
l
t
aneou
sl
y
r
edu
ci
n
g
po
r
t
f
o
li
o
ris
k
,
a
s i
nd
ic
a
t
ed
b
y
t
he
h
ig
he
r S
ha
r
pe
r
a
t
i
o
and
l
o
w
e
r
annua
li
z
ed
v
o
l
a
t
ili
t
y
.
V. C
ON
CL
U
SI
ON
A
N
D
F
U
T
U
RE
W
O
RK
W
e
ha
v
e
i
n
t
r
odu
c
ed
an
i
nno
v
a
t
i
v
e
app
r
oa
c
h
t
o
f
i
nan
ci
a
l s
en
-
t
im
en
t
ana
l
y
sis w
h
ic
h
r
e
s
t
s
upon
t
he
f
i
ne
-
t
un
i
n
g
o
f
a
g
ene
r
a
l
-
pu
r
po
s
e
LLM. I
n
t
h
is w
a
y
,
t
he
p
r
opo
s
ed
m
e
t
hod
ha
s c
ap
i
-
t
a
lis
ed
on
t
he
e
x
t
en
si
v
e
k
no
wl
ed
g
e
ba
s
e
and
r
ea
s
on
i
n
g
ab
ili
t
i
e
s
i
nhe
r
en
t
t
o
LLMs, w
h
ils
t
s
h
i
f
t
i
n
g
t
he
ir
p
rim
a
r
y
ob
j
e
c
t
i
v
e
f
r
o
m
4
Fig. 3: C
o
m
pa
ris
on
o
f
pe
r
f
o
rm
an
c
e
o
f
t
he
35
%
l
on
g
-
s
ho
r
t
po
r
t
f
o
li
o
s w
h
ic
h
w
e
r
e
c
on
s
t
r
u
c
t
ed
u
si
n
g
t
he
f
i
v
e
s
en
t
im
en
t
ana
l
y
sis
m
e
t
hod
s,
f
o
r
t
he
t
im
e
pe
ri
od
o
f
F
eb
r
ua
r
y
2015
t
o
J
une
2021. MA
(
30
)
and
MST
D
(
30
)
r
ep
r
e
s
en
t
t
he
m
o
v
i
n
g
a
v
e
r
a
g
e
and
m
o
v
i
n
g
s
t
anda
r
d
de
v
i
a
t
i
on
, r
e
s
pe
c
t
i
v
e
l
y
,
o
f
t
he
r
e
t
u
r
n
s c
a
lc
u
l
a
t
ed
o
v
e
r
a
30
-
da
y
r
o
lli
n
g wi
ndo
w.
Cumul
a
t
i
v
e
R
e
t
urns
(
%
)
Annu
a
li
z
e
d R
e
t
urn
(
%
)
Sh
a
rp
e
R
a
t
io
Annu
a
li
z
e
d Vol
a
t
ili
t
y
(
%
)
L
M
D
HI
V
-
4VA
D
ER
204.6 100.4 130.6
29.1 13.5 17.9
1.5 0.7 0.9
19.5 18.9 19.6
Fi
n
BERT
213.0
30.3
1.5
20.3
Fi
n
Llama
(
O
u
rs
)
S&
P
500
308.2 83.1
45.0 11.3
2.4 0.62
18.6 18.5
TABLE I: S
t
a
t
is
t
ic
a
l c
o
m
pa
ris
on
be
t
w
een
t
he
f
i
v
e
c
on
si
de
r
ed
s
en
t
im
en
t
ana
l
y
sis m
e
t
hod
s
u
si
n
g
a
35
%
l
on
g
-
s
ho
r
t
po
r
t
f
o
li
o
.
F
o
r C
u
m
u
l
a
t
i
v
e
R
e
t
u
r
n
s, A
nnua
li
z
ed
R
e
t
u
r
n
and
S
ha
r
pe
R
a
t
i
o
,
h
ig
he
r is
be
tt
e
r. F
o
r A
nnua
li
z
ed
V
o
l
a
t
ili
t
y
, l
o
w
e
r is
be
tt
e
r.
t
e
x
t
g
ene
r
a
t
i
on
t
o
cl
a
ssi
f
ic
a
t
i
on
t
a
s
k
s. I
n
add
i
t
i
on
, s
u
c
h
an
app
r
oa
c
h
ha
s
enab
l
ed
t
he
LLMs
t
o
be
c
o
m
e
m
o
r
e
a
tt
uned
t
o
t
he
nuan
c
ed
l
an
g
ua
g
e
o
f
t
he
f
i
nan
c
e
s
e
c
t
o
r, w
h
ils
t
mi
n
imisi
n
g
t
he
ir
r
e
s
ou
rc
e
u
t
ilis
a
t
i
on
and
c
o
m
pu
t
a
t
i
ona
l
de
m
and
s.
O
u
r
f
i
ne
-
t
uned
Ll
a
m
a
2 7B m
ode
l,
t
e
rm
ed
Fi
n
Ll
a
m
a
,
ha
s
been
u
s
ed
t
o
c
on
s
t
r
u
c
t
a
po
r
t
f
o
li
o
,
y
i
e
l
d
i
n
g r
e
s
u
l
t
s
t
ha
t
ha
v
e
s
u
r
pa
ss
ed
t
ho
s
e
o
f
t
he
l
ead
i
n
g c
u
rr
en
t
m
e
t
hod
i
n
t
he
f
i
e
l
d
,
t
he
Fi
n
BER
T
. T
he
Fi
n
Ll
a
m
a
ha
s
a
c
h
i
e
v
ed
c
u
m
u
l
a
t
i
v
e
r
e
t
u
r
n
s w
h
ic
h
ha
v
e
ou
t
pe
r
f
o
rm
ed
t
he
Fi
n
BERT m
ode
l
b
y
44.7
%
, w
h
il
e
a
c
h
i
e
v
i
n
g
a
sig
n
i
f
ic
an
t
l
y
h
ig
he
r S
ha
r
pe
r
a
t
i
o
and
l
o
w
e
r
annua
li
z
ed
v
o
l
a
t
ili
t
y
. T
h
is r
ep
r
e
s
en
t
s
bo
t
h
a
s
ub
s
t
an
t
i
a
l c
on
t
ri
-
bu
t
i
on
t
o
t
he
f
i
e
l
d
o
f
a
c
on
j
o
i
n
t
f
r
a
m
e
w
o
r
k
i
n
v
o
l
v
i
n
g s
en
t
im
en
t
ana
l
y
sis
and
LLMs,
and
de
m
on
s
t
r
a
t
e
s
t
ha
t
f
i
ne
-
t
un
i
n
g
an
LLM
c
an
y
i
e
l
d
s
upe
ri
o
r r
e
s
u
l
t
s,
e
v
en
wi
t
h
a
sm
a
ll
a
m
oun
t
o
f
t
a
s
k-
s
pe
ci
f
ic
da
t
a
. I
n
add
i
t
i
on
,
t
he
p
r
e
s
en
t
w
o
r
k
ha
s s
e
t
a
ne
w
ben
c
h
m
a
r
k
i
n
t
he
f
i
e
l
d
,
t
r
an
sc
end
i
n
g
t
r
ad
i
t
i
ona
l m
ea
s
u
r
e
s s
u
c
h
a
s
a
cc
u
r
a
c
y
and
F1
-
sc
o
r
e
, c
o
mm
on
l
y
u
s
ed
i
n
t
he
li
t
e
r
a
t
u
r
e
.
I
n
s
t
ead
, w
e
f
o
c
u
s
on
p
r
a
c
t
ic
a
l,
f
i
nan
c
e
-
s
pe
ci
f
ic m
e
t
rics w
h
ic
h
ha
v
e
gr
ea
t
e
r r
e
l
e
v
an
c
e
t
o
end
-
u
s
e
rs. I
t
is
ou
r
hope
t
ha
t
s
u
c
h
an
app
r
oa
c
h
is
a
s
t
ep
t
o
w
a
r
d
s
na
rr
o
wi
n
g
do
w
n
t
he
d
i
v
i
de
be
t
w
een
a
c
ade
mic r
e
s
ea
rc
h
and
p
r
a
c
t
ic
a
l
app
lic
a
t
i
on
s wi
t
h
i
n
quan
t
i
t
a
t
i
v
e
f
i
nan
c
e
.
O
u
r
f
u
t
u
r
e
r
e
s
ea
rc
h
will
a
im
t
o
enhan
c
e
bo
t
h
t
he
s
en
t
im
en
t
cl
a
ssi
f
ic
a
t
i
on
a
cc
u
r
a
c
y
and
e
f
f
ici
en
c
y
o
f
ou
r m
ode
l
b
y
i
n
c
o
r
-
po
r
a
t
i
n
g
add
i
t
i
ona
l
t
e
c
hn
i
que
s
t
o
p
r
odu
c
e
an
ea
sil
y
t
r
a
c
t
ab
l
e
p
l
a
t
f
o
rm
t
o
f
a
cili
t
a
t
e
t
he
app
lic
a
t
i
on
o
f
a
r
t
i
f
ici
a
l i
n
t
e
llig
en
c
e
(
AI
)
i
n
t
he
f
i
nan
c
e
s
e
c
t
o
r.
D
isclaim
e
r:
N
o
thing h
e
r
e
in is
f
inancial ad
v
ic
e
, and
N
O
T
a r
e
c
o
mm
e
ndati
o
n t
o
trad
e
r
e
al m
o
n
e
y
.
P
l
e
as
e
us
e
c
o
mm
o
n
s
e
ns
e
and alwa
y
s
f
irst c
o
nsult a pr
o
f
e
ssi
o
nal b
e
f
o
r
e
trading
5
o
r in
v
e
sting.
R
EFERE
N
CES
¨
[
1
]
K. Mis
he
v
, A.
G
j
o
rg
j
e
v
i
k
j
, I. V
oden
s
k
a
, L. T. C
h
i
t
k
u
s
he
v
,
and
D
. Tr
a
-
j
ano
v
, “E
v
a
l
ua
t
i
on
o
f
s
en
t
im
en
t
ana
l
y
sis i
n
f
i
nan
c
e
: Fr
o
m l
e
x
ic
on
s
t
o
t
r
an
s
f
o
rm
e
rs,” IEEE Acc
e
ss,
v
o
l. 8,
pp
. 131662–131682, 07 2020.
[
2
]
H
. T
ou
v
r
on
, L. M
a
r
t
i
n
, K. S
t
one
,
P
. Al
be
r
t
, A. Alm
aha
iri, Y. B
abae
i,
N
. B
a
s
h
l
yk
o
v
, S. B
a
t
r
a
,
P
. B
ha
rg
a
v
a
, S. B
ho
s
a
l
e
e
t
al., “Ll
a
m
a
2:
O
pen
f
ounda
t
i
on
and
f
i
ne
-
t
uned
c
ha
t
m
ode
ls,” a
r
X
iv
p
r
ep
r
i
n
t
a
r
X
iv:2307.09288, 2023.
[
3
]
E. J.
H
u
, Y. S
hen
,
P
. W
a
llis, Z. All
en
-
Z
hu
, Y. Li, S. W
an
g, L. W
an
g,
and
W. C
hen
, “L
o
RA: L
o
w
-
r
an
k
adap
t
a
t
i
on
o
f
l
a
rg
e
l
an
g
ua
g
e
m
ode
ls,” a
r
X
iv
p
r
ep
r
i
n
t
a
r
X
iv:2106.09685, 2021.
[
4
]
E. F. F
a
m
a
, “E
f
f
ici
en
t
c
ap
i
t
a
l m
a
r
k
e
t
s: A r
e
v
i
e
w
o
f
t
heo
r
y
and
e
m
p
iric
a
l
w
o
r
k
,”
T
he
J
ou
r
n
al
of
Fi
n
a
n
c
e
,
v
o
l. 25,
no
. 2,
pp
. 383–417, 1970.
[
5
]
X. Li,
H
. Xi
e
, L. C
hen
, J. W
an
g,
and
X.
D
en
g, “
N
e
ws im
pa
c
t
on
s
t
o
c
k
p
ric
e
r
e
t
u
r
n
v
i
a
s
en
t
im
en
t
ana
l
y
sis,” K
now
l
ed
g
e
-B
as
ed
S
ys
t
em
s,
v
o
l. 69,
pp
. 14–23, 2014.
[
6
]
Z. T. K
e
, B. T. K
e
ll
y
,
and
D
. Xi
u
, “
P
r
ed
ic
t
i
n
g r
e
t
u
r
n
s wi
t
h
t
e
x
t
da
t
a
,”
N
a
t
i
ona
l B
u
r
eau
o
f
Ec
ono
mic R
e
s
ea
rc
h
, T
e
c
h
. R
ep
., 2019.
[
7
]
N
. Cris
t
i
an
i
n
i
and
J. S
ha
w
e
-
T
a
y
l
o
r, A
n
i
n
t
r
odu
c
t
i
on
t
o
s
uppo
r
t
v
e
c
t
o
r
m
ac
h
i
ne
s a
nd
o
t
he
r
k
e
r
ne
l
-
b
as
ed
l
e
a
r
n
i
n
g
me
t
hod
s. C
a
m
b
ri
d
g
e
U
n
i
-
v
e
rsi
t
y
P
r
e
ss, 2000.
[
8
]
Z. Y
an
g,
D
. Y
an
g, C.
D
y
e
r, X.
H
e
, A. Sm
o
l
a
,
and
E.
H
o
vy
, “
H
i
e
r
a
rc
h
ic
a
l
a
tt
en
t
i
on
ne
t
w
o
r
k
s
f
o
r
do
c
u
m
en
t
cl
a
ssi
f
ic
a
t
i
on
,” 01 2016,
pp
. 1480–1489.
[
9
]
J.
D
e
v
li
n
, M.
-
W. C
han
g, K. L
ee
,
and
K. T
ou
t
ano
v
a
, “BERT:
P
r
e
-
t
r
a
i
n
i
n
g
o
f
deep
b
i
d
ir
e
c
t
i
ona
l
t
r
an
s
f
o
rm
e
rs
f
o
r l
an
g
ua
g
e
unde
rs
t
and
i
n
g,” i
n
N
o
r
t
h
A
me
r
ica
n
C
h
a
p
t
e
r
of
t
he
Ass
o
cia
t
i
on
fo
r
C
ompu
t
a
t
i
on
al Li
n
g
u
is
t
ics,
2019.
[O
n
li
ne
]
. A
v
a
il
ab
l
e
:
h
tt
ps://
a
pi.s
e
m
a
n
t
icschol
a
r.or
g
/CorpusI
D
:
52967399
[
10
]
D
. Ar
a
ci, “Fi
n
BERT: Fi
nan
ci
a
l s
en
t
im
en
t
ana
l
y
sis wi
t
h
p
r
e
-
t
r
a
i
ned
l
an
g
ua
g
e
m
ode
ls,” a
r
X
iv
p
r
ep
r
i
n
t
a
r
X
iv:1908.10063, 2019.
[
11
]
Z. C
hen
, S.
G
o
ssi, W. Kim, B. B
e
rm
e
i
t
i
n
g
e
r,
and
S.
H
and
sc
huh
,
“Fi
n
BERT
-
F
O
MC: Fi
ne
-
t
uned
Fi
n
BERT M
ode
l wi
t
h
s
en
t
im
en
t
f
o
c
u
s
m
e
t
hod
f
o
r
enhan
ci
n
g s
en
t
im
en
t
ana
l
y
sis
o
f
F
O
MC mi
nu
t
e
s.”
P
r
o
-
c
eed
i
n
gs
o
f
t
he
4
t
h
ACM I
n
t
e
r
na
t
i
ona
l C
on
f
e
r
en
c
e
on
AI i
n
Fi
nan
c
e
,
2023,
pp
. 357–364.
[
12
]
X.
-
Y. Li
u
,
G
. W
an
g,
and
D
. Z
ha
, “Fi
n
GP
T:
D
e
m
o
cr
a
t
i
z
i
n
g i
n
t
e
r
ne
t
-
sc
a
l
e
da
t
a
f
o
r
f
i
nan
ci
a
l l
a
rg
e
l
an
g
ua
g
e
m
ode
ls,” a
r
X
iv
p
r
ep
r
i
n
t
a
r
X
iv:2307.10485, 2023.
[
13
]
H
. Y
an
g, X.
-
Y. Li
u
,
and
C.
D
. W
an
g, “Fi
n
GP
T:
O
pen
-
s
ou
rc
e
f
i
nan
ci
a
l
l
a
rg
e
l
an
g
ua
g
e
m
ode
ls,” a
r
X
iv
p
r
ep
r
i
n
t
a
r
X
iv:2306.06031, 2023.
[
14
]
B. Z
han
g,
H
. Y
an
g,
and
X.
-
Y. Li
u
, “I
n
s
t
r
u
c
t
-
Fi
n
GP
T: Fi
nan
ci
a
l
s
en
t
im
en
t
ana
l
y
sis
b
y
i
n
s
t
r
u
c
t
i
on
t
un
i
n
g
o
f
g
ene
r
a
l
-
pu
r
po
s
e
l
a
rg
e
l
an
g
ua
g
e
m
ode
ls,” A
r
X
iv,
v
o
l.
ab
s/2306.12659, 2023.
[O
n
li
ne
]
.
A
v
a
il
ab
l
e
:
h
tt
ps://
a
pi.s
e
m
a
n
t
icschol
a
r.or
g
/CorpusI
D
:259224880
[
15
]
S. W
u
,
O
. Irs
o
y
, S. L
u
, V.
D
ab
r
a
v
o
ls
k
i, M.
D
r
ed
z
e
, S.
G
eh
rm
ann
,
P
. K
a
m
badu
r,
D
. R
o
s
enbe
rg,
and
G
. M
ann
, “Bl
oo
m
be
rg
GP
T: A
l
a
rg
e
l
an
g
ua
g
e
m
ode
l
f
o
r
f
i
nan
c
e
,” A
r
X
iv,
v
o
l.
ab
s/2303.17564,
2023.
[O
n
li
ne
]
. A
v
a
il
ab
l
e
:
h
tt
ps://
a
pi.s
e
m
a
n
t
icschol
a
r.or
g
/CorpusI
D
:
257833842
[
16
]
I. L
o
s
h
c
h
il
o
v
and
F.
H
u
tt
e
r, “Fi
x
i
n
g w
e
ig
h
t
de
c
a
y
r
e
g
u
l
a
ri
z
a
t
i
on
i
n
A
da
m,” A
r
X
iv,
v
o
l.
ab
s/1711.05101, 2017.
[O
n
li
ne
]
. A
v
a
il
ab
l
e
:
h
tt
ps://
a
pi.s
e
m
a
n
t
icschol
a
r.or
g
/CorpusI
D
:3312944
[
17
]
T. L
ou
g
h
r
an
and
B. Mc
dona
l
d
, “W
hen
is
a
li
ab
ili
t
y
no
t
a
li
ab
ili
t
y
?
T
e
x
t
ua
l
ana
l
y
sis,
d
ic
t
i
ona
ri
e
s,
and
10
-
Ks,”
T
he
J
ou
r
n
al
of
Fi
n
a
n
c
e
,
v
o
l.
66,
pp
. 35 – 65, 02 2011.
[
18
]
P
. J. S
t
one
,
D
. C.
D
unph
y
, M. S. Smi
t
h
,
and
D
. M.
O
gil
v
i
e
,
T
he
G
ene
r
al
I
nqu
i
r
e
r
: A C
ompu
t
e
r
A
pp
r
o
ac
h
t
o
C
on
t
en
t
A
n
alysis. MIT
P
r
e
ss, 1966.
[
19
]
C.
H
u
tt
o
and
E.
G
il
be
r
t
, “VA
D
ER: A
pa
rsim
on
i
ou
s r
u
l
e
-
ba
s
ed
m
ode
l
f
o
r
s
en
t
im
en
t
ana
l
y
sis
o
f
s
o
ci
a
l m
ed
i
a
t
e
x
t
,”
v
o
l. 08,
no
. 01.
P
r
o
c
eed
i
n
gs
o
f
t
he
8
t
h
I
n
t
e
r
na
t
i
ona
l C
on
f
e
r
en
c
e
on
W
eb
l
o
gs
and
S
o
ci
a
l M
ed
i
a
, ICWSM
2014, 2015,
pp
. 216–225.
[
20
]
Z. T. K
e
, B. T. K
e
ll
y
,
and
D
. Xi
u
, “
P
r
ed
ic
t
i
n
g r
e
t
u
r
n
s wi
t
h
t
e
x
t
da
t
a
,”
N
a
t
i
ona
l B
u
r
eau
o
f
Ec
ono
mic R
e
s
ea
rc
h
, I
n
c,
N
BER W
o
r
k
i
n
g
P
ape
rs
26186, 2019.
[O
n
li
ne
]
. A
v
a
il
ab
l
e
:
h
tt
ps://Econ
P
a
p
e
rs.r
e
p
e
c.or
g
/R
e
P
Ec:
nbr:nb
e
rwo:26186
[
21
]
J. B. B
e
r
k
and
P
. M.
D
e
M
a
r
z
o
, “C
o
r
po
r
a
t
e
f
i
nan
c
e
,”
v
o
l. 5, 2019.
[
22
]
Y
ahoo
Fi
nan
c
e
, “Tr
ea
s
u
r
y
y
i
e
l
d
10
y
ea
rs
h
is
t
o
ric
a
l
da
t
a
.” 2023.
[O
n
li
ne
]
. A
v
a
il
ab
l
e
:
h
tt
ps://
f
in
a
nc
e
.
y
a
hoo.com/quo
t
e
/
%
5ET
N
X/his
t
or
y
VI. A
PP
E
N
D
IX
A. L
e
xic
on
-B
as
ed
A
pp
r
o
ac
he
s
1
)
H
a
r
va
r
d
I
V-
4
P
syc
ho
l
o
gical
D
ic
t
i
on
a
r
y
(
H
I
V-
4
)
:
H
IV
-
4
is
one
o
f
t
he
o
l
de
s
t
m
anua
ll
y
c
on
s
t
r
u
c
t
ed
l
e
x
ic
on
s,
and
is
u
s
ed
f
o
r
ob
j
e
c
t
i
v
e
l
y
i
den
t
i
fy
i
n
g s
pe
ci
f
i
ed
c
ha
r
a
c
t
e
ris
t
ics
o
f
m
e
ss
a
g
e
s i
n
a
r
ea
s i
n
v
o
l
v
i
n
g s
o
ci
a
l sci
en
c
e
,
po
li
t
ic
a
l sci
en
c
e
,
and
p
s
y
c
ho
l
o
g
y
. T
he
l
a
t
e
s
t
v
e
rsi
on
o
f
t
he
H
IV
-
4
d
ic
t
i
ona
r
y
c
on
t
a
i
n
s
o
v
e
r 11,000 w
o
r
d
s w
h
ic
h
a
r
e
cl
a
ssi
f
i
ed
i
n
t
o
one
o
r
m
o
r
e
o
f
183 c
a
t
e
g
o
ri
e
s. I
n
t
h
is w
o
r
k
, w
e
f
o
c
u
s
on
t
he
1,045
w
o
r
d
s l
abe
ll
ed
a
s
po
si
t
i
v
e
and
t
he
1,160 w
o
r
d
s l
abe
ll
ed
a
s
ne
g
a
t
i
v
e
.
2
)
L
ou
g
h
r
a
n
a
nd
Mc
D
on
al
d
(
LM
D
)
D
ic
t
i
on
a
r
y: L
ou
g
h
r
an
and
Mc
D
ona
l
d
e
v
a
l
ua
t
ed
s
t
anda
r
d
d
ic
t
i
ona
ri
e
s
and
f
ound
t
ha
t
t
he
s
e
f
r
equen
t
l
y
miscl
a
ssi
fy
t
e
rms wi
t
h
i
n
f
i
nan
ci
a
l
t
e
x
t
s. T
h
is
i
n
sig
h
t
l
ed
t
o
t
he
de
v
e
l
op
m
en
t
o
f
t
he
LM
D
d
ic
t
i
ona
r
y
, w
h
ic
h
is
s
pe
ci
f
ic
a
ll
y
t
a
il
o
r
ed
f
o
r
t
he
f
i
nan
ci
a
l s
e
c
t
o
r. T
he
d
ic
t
i
ona
r
y
c
a
t
e
g
o
ri
z
e
s w
o
r
d
s i
n
t
o
si
x
d
is
t
i
n
c
t
s
en
t
im
en
t
c
a
t
e
g
o
ri
e
s:
ne
g
a
t
i
v
e
,
po
si
t
i
v
e
,
un
c
e
r
t
a
i
n
t
y
, li
t
igi
ou
s, s
t
r
on
g m
oda
l,
and
w
ea
k
m
oda
l. I
t
w
a
s c
on
s
t
r
u
c
t
ed
u
si
n
g
da
t
a
f
r
o
m 50,115 10
-
K
f
ili
n
gs
f
r
o
m 8,341
f
irms lis
t
ed
on
t
he
N
e
w Y
o
r
k
S
t
o
c
k
E
x
c
han
g
e
(N
YSE
)
and
t
he
N
a
t
i
ona
l Ass
o
ci
a
t
i
on
o
f
S
e
c
u
ri
t
i
e
s
D
ea
l
e
rs A
u
t
o
m
a
t
ed
Q
uo
t
a
t
i
on
s
(N
AS
D
A
Q)
, c
o
v
e
ri
n
g
t
he
pe
ri
od
f
r
o
m 1994
t
o
2008.
Ov
e
r
a
ll,
t
he
LM
D
d
ic
t
i
ona
r
y
c
on
t
a
i
n
s 2,355
ne
g
a
t
i
v
e
f
i
nan
ci
a
l w
o
r
d
s
and
353
po
si
t
i
v
e
f
i
nan
ci
a
l w
o
r
d
s.
3
)
V
al
en
c
e
A
w
a
r
e
D
ic
t
i
on
a
r
y
fo
r
sE
n
t
i
men
t
R
e
as
on
i
n
g
(V
A
D
E
R)
:
V
A
D
ER c
o
m
b
i
ne
s l
e
x
ic
a
l
f
ea
t
u
r
e
s,
de
ri
v
ed
f
r
o
m
micr
o
-
b
l
o
g c
on
t
e
x
t
s, wi
t
h
t
he
gr
a
mm
a
t
ic
a
l
and
s
y
n
t
a
c
t
ic
a
l
c
on
v
en
t
i
on
s
t
ha
t
hu
m
an
s
t
y
p
ic
a
ll
y
e
m
p
l
o
y
t
o
e
x
p
r
e
ss
o
r
e
m
-
pha
si
z
e
s
en
t
im
en
t
i
n
t
en
si
t
y
. T
h
is
enab
l
e
s
V
A
D
ER
t
o
a
cc
u
r
a
t
e
l
y
quan
t
i
fy
t
he
s
en
t
im
en
t
s
t
r
en
g
t
h
o
f
t
e
x
t
. T
he
m
ode
l c
on
t
a
i
n
s
app
r
o
x
im
a
t
e
l
y
9,000
t
o
k
en
f
ea
t
u
r
e
s, w
h
ic
h
a
r
e
ea
c
h
a
ssig
ned
a
s
en
t
im
en
t
sc
o
r
e
r
an
gi
n
g
f
r
o
m
-
4
(
i
nd
ic
a
t
i
n
g
e
x
t
r
e
m
e
l
y
ne
g
a
t
i
v
e
s
en
t
im
en
t
)
t
o
+
4
(
i
nd
ic
a
t
i
n
g
e
x
t
r
e
m
e
l
y
po
si
t
i
v
e
s
en
t
im
en
t
)
. T
he
o
v
e
r
a
ll
po
l
a
ri
t
y
sc
o
r
e
f
o
r
a
t
e
x
t
is c
a
lc
u
l
a
t
ed
b
y
s
u
mmi
n
g
t
he
s
en
t
im
en
t
sc
o
r
e
s
o
f
ea
c
h
w
o
r
d
p
r
e
s
en
t
i
n
t
he
l
e
x
ic
on
, wi
t
h
t
he
f
i
na
l sc
o
r
e
no
rm
a
li
z
ed
t
o
f
a
ll wi
t
h
i
n
t
he
r
an
g
e
o
f
-
1
t
o
+
1.
B
.
D
eep
L
e
a
r
n
i
n
g A
pp
r
o
ac
he
s
1
)
Fi
n
B
E
RT
: Fi
n
BERT l
e
v
e
r
a
g
e
s
t
he
BERT m
ode
l
a
rc
h
i
t
e
c
-
t
u
r
e
,
and
is s
pe
ci
f
ic
a
ll
y
t
a
il
o
r
ed
f
o
r
f
i
nan
ci
a
l c
on
t
e
x
t
s. I
t
w
a
s
p
r
e
-
t
r
a
i
ned
on
a
s
ub
s
t
an
t
i
a
l
f
i
nan
ci
a
l
t
e
x
t
c
o
r
pu
s c
on
sis
t
i
n
g
o
f
1.8M
ne
ws
a
r
t
icl
e
s s
ou
rc
ed
f
r
o
m
t
he
T
ho
ms
on
R
eu
t
e
rs T
e
x
t
R
e
s
ea
rc
h
C
o
ll
e
c
t
i
on
(
TRC2
)
da
t
a
s
e
t
, s
pann
i
n
g
t
he
y
ea
rs
be
-
t
w
een
2008
t
o
2010. F
u
r
t
he
r r
ef
i
ne
m
en
t
w
a
s
a
c
h
i
e
v
ed
t
h
r
ou
g
h
f
i
ne
-
t
un
i
n
g
on
t
he
Fi
nan
ci
a
l
P
h
r
a
s
eban
k
(
F
P
B
)
da
t
a
s
e
t
,
t
hu
s
enhan
ci
n
g i
t
s c
apab
ili
t
i
e
s i
n
f
i
nan
ci
a
l s
en
t
im
en
t
cl
a
ssi
f
ic
a
t
i
on
.
Fi
n
BERT g
ene
r
a
t
e
s S
o
f
t
M
a
x
ou
t
pu
t
s
f
o
r
t
h
r
ee
l
abe
ls:
po
si
t
i
v
e
,
ne
g
a
t
i
v
e
,
and
neu
t
r
a
l.
6